Monitoring for intrusion behaviors with a small UAV

In this paper we present behavior recognition algorithms using images captured by a small UAV (unmanned aerial vehicle). Small UAVs provide a mobile platform capable of monitoring a large area, but each UAV has a limited field of view, resulting in very brief observations of targets of interest as UAVs fly over them. Standard behavior recognition algorithms are designed for CCTV networks, where cameras have a fixed view of the target area. To apply these algorithms to imagery from small UAVs, we first geo-locate targets based on their image coordinates and correspondences between geo-located landmarks and landmarks in each video frame. We then apply behavior recognition algorithms that are suited to very short observation periods. We test these algorithms in a simulation environment incorporating realistic target motion and validate the entire approach on imagery captured by a small UAV. These experiments illustrate that automatic behavior recognition has merit even during brief fly-overs, typical of a small UAV on surveillance patrol.